Ad Keyframe Extraction and Product Detection Framework in Short Video Advertising

Short video advertising attracts users through authentic expressions, reducing resistance and improving conversion rates. However, as traffic costs rise, simply increasing exposure is insufficient. Advertisers need high‐quality ads that resonate with consumers, making visual optimization essential. Accurately identifying and locating products within images is crucial for achieving this. However, short videos often contain redundant frames, making it difficult to extract keyframes representing core content. To address this, we propose an Advertising Keyframe Extraction and Product Detection Framework (AKEPDF), which operates in two stages: first, using an unsupervised K‐Nearest Neighbor algorithm to segment videos and select clear frames as keyframes; then, using EfficientViT‐YOLOv8n for product detection in keyframes. Our experiments demonstrate that AKEPDF effectively extracts keyframes and detects product locations and types, outperforming baseline models.